AI Guides › Step-by-step guides
By Nigel Guy · 7 min read
Claude's memory now carries between ordinary chats and Cowork tasks that run in the cloud, and most people handle it one of two ways. Either they leave it to fill itself from whatever happened to come up, so it remembers a one-off client from March and not the business you run, or they pour their whole life into it and then wonder why a draft for a supplier mentions their divorce. Both feel fine because memory works quietly. Both mean Claude starts each session knowing the wrong things.
The rule: memory holds the durable facts about you and how you work; anything that belongs to one piece of work goes in that project's instructions, and anything you would not say to a new colleague stays out.
| You need | Detail at time of writing |
|---|---|
| A Claude account | Memory is on by default on Free, Pro and Max. On Team and Enterprise an owner has to switch it on first, in Organization settings > Capabilities. |
| Web or Claude Desktop | Memory import works on web and Desktop, on Free, Pro, Max and Team plans. |
| Cowork in the cloud (optional) | Shared memory only reaches Cowork when the task runs in the cloud. Cowork sessions that run locally on your computer do not use it. |
| Cost | Nothing extra: memory is included on every plan. Claude's pricing page lists Pro at $20 a month in US dollars (about £15 at time of writing); the sterling price you are charged depends on your region and VAT, so check the pricing page shown at checkout. |
| 30 minutes | Twenty to write the brief, ten to check what Claude saved. |
Not available: organisations on HIPAA, public-sector or custom data-retention agreements.
Open Settings > Memory. Confirm Generate memory from chats is on. Leave Include sensitive topics in memory off unless you have a specific reason, such as wanting dietary needs remembered for meal planning. With it off, Claude does not store health, religious beliefs, ethnicity, political views or gender identity. Some things are never stored even with it on: government ID numbers, financial account numbers, criminal history and immigration status.
While you are there, look at the existing entries under Topics. If you used Claude before the August 2026 change, this is what it already believes about you.
Before writing anything, split your context. This is the step most people skip, and it is where the dumping happens.
| Goes in personal memory | Goes in a project's instructions | Stays out entirely |
|---|---|---|
| Your name, location, what the business does in a line | The brief, audience and tone for one client or product | Passwords, API keys, account numbers |
| Your role and how you want Claude to work with you | Reference files and style guides (as project knowledge) | Other people's personal details beyond name and role |
| Tools you actually use and where files live | Rules that only apply to that work | Health, family or legal matters you would not tell a contractor |
| Answer length, format, British English | Deadlines and numbers for that project | Anything confidential to a client |
| Your hard "ask me first" rules | Plans you have not decided on yet | |
| The few people and projects that recur every week |
The test for the middle column: would this fact be wrong, or a leak, in a chat about something else? If yes, it belongs in a project. Each project has its own separate memory space, so project facts stay there.
Start a fresh chat outside any project and paste the prompt below. Fill in the bracketed items you know; leave the rest and Claude will ask.
You are helping me write a short memory brief that Claude will keep about me across all future chats. It must hold only durable facts about me and how I like to work, not details of any single project.
My raw notes: [PASTE_ANYTHING_YOU_HAVE — name, location, business, tools, preferences, recurring people]
Steps:
1. Read my notes. List any of these sections you cannot fill, and ask me about them, one short question each. Do not invent answers.
2. Once I reply, write the brief under these headings:
- About me: name, base, what I do in one sentence.
- How we work: my role versus yours (for example, I decide, you draft and challenge).
- Current focus: the single priority this quarter and its target, with a number and a date.
- Tools and files: only tools I said I use daily, and where my work lives.
- How to answer: default length, preferred format, spelling (British English), and that you should disagree when I am wrong rather than soften it.
- Ask first: actions you must check with me before doing.
- Regulars: up to five people, clients or projects, one line each, names and roles only.
3. Keep it under 250 words, plain bullet points, no adjectives about me.
Rules: leave out passwords, account numbers, health, family, legal or client-confidential details even if they appear in my notes, and tell me what you left out. If a fact only matters to one project, move it to a separate list headed "Put in a project instead".
Before you reply, check: every line is something you would still want to know in six months; nothing is guessed; the word count is under 250.
Fill in: your raw notes. Expect a round of questions before the brief appears.
Read the result and cut anything that is true this week but not next quarter.
Go to Settings > Memory and choose Start import. Paste the brief into the box and click Add to memory. When it finishes, click See what Claude learned about you.
Anthropic describes import as experimental: Claude may not take in everything, and it favours work-related topics over personal ones. That is why you check in Step 5 rather than assuming. If the import misses a line, you can also say "Remember that..." in an ordinary chat.
Put the "Put in a project instead" list into the relevant project: open the project, click Set project instructions, paste, and click Save instructions.
Back in Settings > Memory, open each entry under Topics. Use the edit icon to correct wording, and delete anything stale or wrong. Edits apply to all future conversations. Do this once now, then monthly.
For a quick audit from inside a chat, ask Claude: "Write out your memories of me verbatim, exactly as they appear in your memory." Save the output as a dated text file; that is also your backup.
For a chat that should neither read nor add to memory, click + in the message box before your first message and switch Memory off. For anything you do not want in your history at all, use an incognito chat (the ghost icon when starting a new chat outside a project). Use these for personal admin, a sensitive HR question, or testing how Claude answers a stranger.